Meta’s Muse Wonder went on sale through the Meta Model API on October 12, 2026, and Muse Wonder pricing matches the list price of Claude Opus 5.5. The catch is in which model you get: the higher-scoring version, Muse Wonder Full, is not on the API at all.
- Muse Wonder costs US$4 per million input tokens, US$0.20 cached and US$20 output, with a 1M-token context window.
- Muse Wonder Full is available only inside Muse Code, on the Power Usage plan ($50 a month) and Maximum Usage plan ($200 a month).
- Meta’s full safety report is listed as coming “soon,” so the evidence a security review would ask for is not published yet.
When a lab releases two sizes of a model at once, the larger one usually gets the charts and the smaller one gets the price list. Meta followed that pattern. Its research post introduces Muse Wonder and Muse Wonder Full as its “most capable models” for personal and professional agent work, and its developer page shows which one a Canadian team can actually call from its own software.
What Meta released, and where
Two models. Muse Wonder is live in the Meta Model API and in Muse Code, Meta’s command-line coding tool. Muse Wonder Full is live only in Muse Code, on the two most expensive plans, and has no API price. Meta says Muse Wonder “will soon power experiences” in its Muse app, rolling out to subscribers first.
Meta describes two features that matter for business use. Active Memory is “carrying important context about you across conversations.” Active Focus is how the model keeps working through blockers, and Meta says that “along the way, it asks for clarification when needed and checks in before taking important actions.” We have not confirmed which of these features are available to API customers in Canada.
What the pricing compares to
At US$4 input and US$20 output per million tokens, Muse Wonder matches the list price of Claude Opus 5.5 exactly. It costs twice Claude Sonnet 5.5, which Anthropic lists at US$2 and US$10. Cached input is US$0.20 per million tokens, a twentieth of the regular input rate.
Equal list prices do not mean equal bills. Cost per task depends on how many tokens each model uses to finish the same job, and nobody outside Meta has published that comparison for business workloads yet.
What Meta measured, and who chose the tests
Meta’s comparison table mixes its own benchmarks, including an internal “Personal Superintelligence Index,” with outside ones. The most useful row for a buyer is the Artificial Analysis Intelligence Index, an independent composite, which Meta reports at maximum reasoning effort.
| Model (max effort) | Artificial Analysis Index, as reported by Meta | Available on the API? |
|---|---|---|
| Claude Opus 5.5 | 58 | Yes (Anthropic) |
| Muse Wonder Full | 56 | No, Muse Code only |
| Claude Fable 5.1 | 53 | Not checked |
| GPT-6 Astra | 53 | Not checked |
| Muse Wonder | 52 | Yes (Meta Model API) |
On Meta’s own figures, the model a developer can buy through the API scores six points below Opus 5.5 on that index at the same list price. Full narrows the gap to two points, but you can only use it through Meta’s coding tool. Meta’s internal benchmarks show its models ahead on several agent and memory tests; those are tests Meta designed, and we treat them as claims.
What we ran and what we didn’t
We read Meta’s research post, its newsroom announcement and its developer pricing page. We did not run Muse Wonder or Muse Wonder Full, and we did not reproduce any benchmark.
The Limitation
Nobody outside Meta can yet check how Muse Wonder handles private information. Meta says it trained the family “to avoid sharing private information without permission,” and that “a full report with additional safety evaluations will be available soon.” For a model sold on remembering context across conversations, that report is the document a privacy or security review would start with. Its publication, with test methods, would remove this limit.
What would change our mind
Two things. Muse Wonder Full arriving on the API with a published price would make the table above a fair comparison. Independent cost-per-task results on business tasks, rather than coding, would show whether equal list prices produce equal bills.
A buyer’s test
If you already pay for Opus 5.5, pick 20 real tasks from last month that it handled well. Run them on Muse Wonder through the API at the same effort level, record tokens used and minutes of checking per result, and have the person who owns the work grade each output pass or fail. Pass if Muse Wonder matches at least 18 of 20 at equal or lower total cost. Two afternoons. Keep work data out of any memory feature until the safety report is published. Anthropic’s own price history is a useful check, since Claude Opus 4.8 pricing held flat while bills still moved.
More model reviews are in AI Tools and Reviews.
FAQ
How much does Muse Wonder cost?
Meta lists Muse Wonder at US$4 per million input tokens, US$0.20 per million cached input tokens and US$20 per million output tokens on the Meta Model API.
Can I use Muse Wonder Full through the API?
No. At launch, Meta says Muse Wonder Full is available only in Muse Code on the Power Usage and Maximum Usage plans, listed at $50 and $200 a month.
How does Muse Wonder compare with Claude Opus 5.5?
The list prices are identical. On the Artificial Analysis Index as reported by Meta, Muse Wonder scores 52 and Opus 5.5 scores 58 at maximum effort.
Has Meta published a safety report for Muse Wonder?
Meta says a full report with additional safety evaluations will be available soon, so none was published at launch.
